Variational Autoencoder in Conditional Generative Adversarial Neural Networks Model

Generative models make up a significant part in unsupervised learning problems. Generative models can also be conditioned on the additional data information such as class labels. Conditional generative models can, for example, be used in image super-resolution or inpainting tasks. Another promising application for such models is a dataset augmentation problem, so a good generative model can improve generalization ability of a supervised classifier. The most known and used models in this field are Generative Adversarial Networks (GAN) and Variational Autoencoder (VAE) and each has its own disadvantages. Nowadays, an active research is carried out in order to combine these two models into single model which will balance these two approaches. In the proposed method, VAE is combined with more advanced GAN model, the Wasserstein GAN, in order to be able to solve hidden image representation inference problem and to make training process more stable. This model generates visually correct images, and has higher evaluation metrics scores, compared to VAE and GAN.